1.03pw/b ultra-low leakage voltage-stacked sram for
TRANSCRIPT
1.03pW/b Ultra-low Leakage Voltage-Stacked SRAM for Intelligent Edge Processors Jingcheng Wang, Hyochan An, Qirui Zhang, Hun Seok Kim, David Blaauw, Dennis Sylvester
University of Michigan, Ann Arbor (email: [email protected])Abstract
A stacked voltage domain SRAM is proposed where arrays are split into two sets (top and bottom) with their supplies connected in series. System supply current is reused by top and bottom sets, and supply voltage is divided among the two sets of arrays, enabling seamless integration of very low voltage SRAM retention in a larger system with a nominal supply, without need for an efficiency-reducing LDO. An array swapping approach provides stable access to arbitrary banks within one system clock cycle. A comprehensive sizing strategy (W&L) is employed to optimally balance hold stability and bitcell size. Integrated in an IoT imaging system in 40nm CMOS, the proposed 8.9Mb SRAM achieves 1.03pW/bit leakage, a >100× reduction over conventional SRAM in the same technology.
Keywords: low leakage, sub/near threshold, SRAM, voltage stacking, charge recycling, array swapping
Introduction Intelligent IoT devices seek to fit complete neural network
models into on-chip memories to avoid costly off-chip DRAM accesses. As a result, SRAMs can consume >80% of chip area and 90% of standby power [1]. Prior work focused on reducing SRAM leakage via techniques such as HVT/thick-oxide devices [2-3], reverse body bias [4-5], floating bitline [6], raising VSS [7-8], and lowering VDD [9-10]. Apart from HVT usage, which enables 10× leakage reduction and is readily deployed, supply voltage lowering is one of the most effective approaches to reduce leakage due to the DIBL effect. However, it raises two issues: 1) Commercial bitcells are not sized to hold data at subthreshold supplies and require a careful hold margin / density tradeoff; 2) Voltage regulation is required to generate the voltage level for SRAM arrays. LDOs are conventionally used, incurring area and power overheads due to efficiency loss. Voltage stacking generates an intermediate voltage level by placing voltage domains in series and has been used in microprocessors [11] and high bandwidth data buses [12]. The major challenge in voltage stacking is balancing the active current between top and bottom levels, to maintain a stable mid-rail voltage level. This often requires an additional voltage regulator, reducing the stacking benefits. SRAM arrays, however, are dominated by near-constant leakage (writing a bit draws 10s of pA average current versus µA-level background leakage), making them ideal for voltage stacking.
Voltage Stacking and Array Swapping To allow access to arbitrary arrays during operation while
avoiding insertion of complex level converters, we propose a novel array voltage swapping mechanism. The SRAM peripherals are not stacked and therefore wordline and bitline voltages remain at VDDcore (~2VDDmid) for faster operation speed, inherent write/read noise margin enhancement (Fig. 1). This also removes the need for level converters but, as a result, only bottom arrays can be read/written. When a top array is accessed, the memory controller first swaps the voltage of a bottom array in the same quad-array SRAM bank with the desired top array (Fig. 3). This swap mechanism ensures the leakage current remains balanced and is completed in one system clock cycle due to the relatively low IoT processer clock frequency. In addition to leakage reduction from reduced supply voltage, the approach offers an additional 2× leakage reduction in top arrays due to their inherent reverse body bias and reduced bitline leakage effects. As a result, total leakage is minimized by increasing the fraction of top arrays to > 50% (e.g., 75%); this is supported by measurements while the optimal ratio can be set by a memory controller.
Memory Cell and Super-Cutoff Read Fig. 2 shows the bitcell schematic and layout. The cross-coupled
4T uses HVT devices to minimize hold leakage while LVT devices in the read port provide faster sensing speed. The bitcell is upsized for improved hold noise margin (HNM). Channel length is increased to the point where leakage is minimum (18% less), also improving HNM by 10% while incurring 8% cell density loss. Channel width is increased, initially improving HNM faster than leakage power, providing a favorable tradeoff. Final sizes are chosen to balance among density, HNM, and leakage. To decouple read/write operation, we use a Z8T structure [13] instead of a traditional 8T, as differential sensing provides faster read speed and larger sensing margin. Since in our stacked SRAM the array voltage is ~½ the bitline voltage, it inherently avoids the clamping current problem in the original Z8T as unselected cells are super-cutoff with negative VGS. Further, the stacked configuration results in word-line overdrive, greatly increasing write margin.
Bank Architecture and Swapping Control Each SRAM bank has 4 arrays with power switches that connect
an array to either top or bottom voltage domains (Fig. 3). Each bank can have 0−3 top arrays but at least one array must be in the bottom domain. When accessing a top array, the SRAM controller swaps this array with a bottom array in the same bank in two steps: First, the two arrays (target and swapping) are expanded to full voltage, after which they are collapsed to the appropriate half range. Since the two arrays are physically close, local charge sharing minimizes the disturbance to the mid-rail. All on-chip SRAM arrays in the system are connected to the same power/ground/mid-rail, resulting in a large amount of innate decoupling capacitance and background load current to suppress transient noise. To smooth transitions and reduce coupling noise, each power switch consists of both small and large headers/footers that are turned on in sequence. Each swap consumes ~8pJ, which is comparable to a single 128b read. To minimize the frequency of swaps, instruction memories (exhibiting mainly random accesses) are placed in the bottom domain, whereas neural engine memories with mostly sequential access patterns are primarily placed in the top domain. SRAM peripherals are power gated immediately after each access to reduce leakage.
Measurement Results The proposed SRAM was implemented in a 40nm CMOS image
processing IoT chip with 8.9Mb memories (Fig. 4). Fig. 5 shows measured leakage across voltage and temperature. As the number of top arrays increases, the mid-rail voltage rises while leakage continually drops (Fig. 6). Fig. 7 shows excellent mid-rail voltage stability; VDDmid varies only ±16mV across 100°C, drops ≤1.74mV when arrays swap every 11 cycles, and is unaffected with read/writes every cycle at full speed. The design operates at 438kHz at 0.7V (enabling 14fps in the supported image processing system) with 67fJ/bit access energy at 0.6V. Fig. 8 compares this work to other state-of-the-art low leakage SRAMs. The proposed work achieves 1.03pW/b at 0.58V without using large thick-oxide device, extra supply levels, or SOI with aggressive body biasing.
References [1] S. Han et al., ISCA 2016. [2] G. Chen et al., ISSCC 2010 [3] T. Fukuda et al., ISSCC 2014. [4] R. Ranica et al., VLSI 2013 [5] M. Yabuuchi et al., VLSI 2017. [6] Y. Wang et al., ISSCC 2009 [7] J. Chang et al., JSSC 2007. [8] T. Kim et al., JSSC 2009 [9] F. Hamzaoglu et al., ISSCC 2008. [10] N. Verma et al., JSSC 2008 [11] K. Blutman et al., JSSC 2017. [12] J. Wilson et al., ISSCC 2016 [13] J. Wu et al., JSSC 2011. [14] Q. Dong et al., VLSI 2017
978-1-7281-9942-9/20/$31.00 ©2020 IEEE
Authorized licensed use limited to: University of Michigan Library. Downloaded on August 19,2020 at 17:56:50 UTC from IEEE Xplore. Restrictions apply.
[15] A. Teman et al., JSSC 2011. [16] N. Maeda et al., JSSC 2013 [17] F. Tachibana et al., ISSCC 2013. [18] F. Frustaci et al., ISSCC 2017. [19] H. Fujiwara et al., VLSI 2013
HVT Stacking:Vdd
scaling&sizing Stacking:
top array Vsb
&bitline effects
7x
11x
2x
2x
VDDmid=½VDDcore
VDDcore=0.6V, 75% Top arrays
VDDcore=0.6V
NEURAL ENGINE(6Mb SRAM)
IMAGE COMP MEM 1.9Mb
INSTR MEM0.5Mb
CD MEM 0.5Mb
ArrayArrayArray
VDDcore 0.6V
VSS 0V
VDDarray 0.3V
SRAM
Per
iph.
&
Chip
Log
ic
SRAM
Per
iph.
&
Chi
p Lo
gic
LDO
VDDcore 0.6V
VSS 0V
VDDmid 0.3V
SwapWhenTargetArray
On Top
Conventional approach: LDO Proposed approach: Voltage Stacking
11x
0.60.3
Isub
Isub
0.3Isub
0
Isub0.3Isub
0
Isub
SRAM Arrays
Top Arrays
Bottom Arrays
NEURAL ENGINE(6Mb SRAM)
IMAGE COMP MEM 1.9Mb
INSTR MEM0.5Mb
CD MEM 0.5Mb
ArrayArrayArray
VDDcore 0.6V
VSS 0V
VDDarray 0.3V
SRAM
Per
iph.
&
Chip
Log
ic
SRAM
Per
iph.
&
Chi
p Lo
gic
LDO
VDDcore 0.6V
VSS 0V
VDDmid 0.3V
SwapWhenTargetArray
On Top
Conventional approach: LDO Proposed approach: Voltage Stacking
11x
0.60.3
Isub
Isub
0.3Isub
0
Isub0.3Isub
0
Isub
SRAM Arrays
Top Arrays
Bottom Arrays
Fig. 1 Proposed array stacking and swapping technique.
Neural Engine Memory(6Mb)
Instruction Memory(0.5Mb)
CD Memory (0.5Mb)
Compressed Image Memory (1.9Mb)
5.6mm
4.8m
m
Fig. 4 Die photograph and shmoo plot.
This work
[19][18]
[17]
[4]
[3][5]
[16]
[2]
[10][15]
[14] [13][8]
This work
[19][18]
[17]
[4]
[3][5]
[16]
[2]
[10][15]
[14] [13][8]
Fig. 8 Comparison table and design space landscape.
HVT Stacking:Vdd
scaling&sizing Stacking:
top array Vsb
&bitline effects
7x
11x
2x
2x
VDDmid=½VDDcore
VDDcore=0.6V, 75% Top arrays
VDDcore=0.6V
HVT Stacking:Vdd
scaling&sizing Stacking:
top array Vsb
&bitline effects
7x
11x
2x
2x
VDDmid=½VDDcore
VDDcore=0.6V, 75% Top arrays
VDDcore=0.6V
Top Array 0.6-0.3V
Power Gate Peripheral
Bottom Array 0.3-0V
Always On 0.6VVoltage Domain:
Chip Sleep Domain
ISOL
ATIO
N ISOLATION
Chip Computing
Logic
Chip Sleep EN
SRAM Bank
EN
×16
ISOL
ATIO
N ISOLATION
VSS VDD VSS VDD
VSS VDD VSS VDD
VDD
VSS
SRAM
Con
trolle
r
LAT
LATPG_EN
VSS
VDD
ADDR, DINDOUT
CLK Gate
VDDcore
VSS VDDmid
VDD
VSSSystem
CLK SRAM CLK Deco
der/W
L Dr
iver
Column Peripheral / Timing Control/ IO
Deco
der/W
L Dr
iver
128 x 128 BitcellArray3
128 x 128 BitcellArray4
128 x 128 BitcellArray1
128 x 128 BitcellArray2
Fig. 2 Bitcell schematic and layout (top), hold noise margin (HNM) and leakage versus bitcell sizing (bottom-left), currents on the bitline during read operation (bottom-right).
Fig. 3 SRAM bank architecture (top), array swapping algorithm and power switches (bottom).
Top Array 0.6-0.3V
Power Gate Peripheral
Bottom Array 0.3-0V
Always On 0.6VVoltage Domain:
Chip Sleep Domain
ISOL
ATIO
N ISOLATION
Chip Computing
Logic
Chip Sleep EN
SRAM Bank
EN
×16
ISOL
ATIO
N ISOLATION
VSS VDD VSS VDD
VSS VDD VSS VDD
VDD
VSS
SRAM
Con
trolle
r
LAT
LATPG_EN
VSS
VDD
ADDR, DINDOUT
CLK Gate
VDDcore
VSS VDDmid
VDD
VSSSystem
CLK SRAM CLK Deco
der/W
L Dr
iver
Column Peripheral / Timing Control/ IO
Deco
der/W
L Dr
iver
128 x 128 BitcellArray3
128 x 128 BitcellArray4
128 x 128 BitcellArray1
128 x 128 BitcellArray2
Fig. 7 Mid-rail variation with temperature (left), voltage drop due to various memory activities (right).
Fig. 5 Leakage across temperature and voltage.
Fig. 6 Mid-rail voltage and leakage with number of top arrays.
Acknowledgement We thank Sony Semiconductor Solutions Corp./Sony Electronics Inc. for supporting this work
Authorized licensed use limited to: University of Michigan Library. Downloaded on August 19,2020 at 17:56:50 UTC from IEEE Xplore. Restrictions apply.